Whisper 1: Complete Specifications, Pricing, API Access & Use Cases (2026)
What is Whisper 1?
Whisper 1 is OpenAI’s general-purpose automatic speech-recognition model, originating from the Whisper system released on September 21, 2022, with multilingual transcription, language identification, speech-to-English translation, and hosted API pricing of $ 0.006 per audio minute as of July 2026.
OpenAI describes Whisper as a Transformer-based speech-recognition system trained on 680,000 hours of multilingual and multitask supervised audio collected from the web. Its training design supports multilingual speech recognition, language identification, timestamp prediction, and translation from supported languages into English.
The hosted OpenAI API model uses the model ID whisper-1 . It accepts audio input and returns text output; it is not a conversational large language model, reasoning model, image model, or text-to-speech system. OpenAI classifies it as a general-purpose speech-recognition model with audio input and text output.
Whisper 1 is commonly evaluated for interview transcription, meeting records, lecture notes, subtitles, media indexing, searchable audio archives, and workflows that translate recorded speech into English.
What Are Whisper 1’s Key Specifications and Pricing?
| Specification | Verified Value |
|---|---|
| Provider | OpenAI (as of July 2026) |
| Model Family | Whisper (as of July 2026) |
| Model Type | General-purpose automatic speech-recognition model (as of July 2026) |
| Original Release Date | September 21, 2022 (as of July 2026) |
| Hosted API Model ID | whisper-1 (as of July 2026) |
| Gate.AIModel ID | openai/stt-whisper-1, as per theGate.AImodel-card athttps://gate.ai/models(as of July 2026) |
| Context Window | Not specified as a conventional token context window in official documentation (as of July 2026) |
| OpenAI Input Pricing | \$0.006 per audio minute (as of July 2026) |
| Cached Input Pricing | Not applicable or not specified (as of July 2026) |
| Output Pricing | No separate text-output price is specified; usage is billed by audio duration (as of July 2026) |
| Pricing Unit | Audio minute (as of July 2026) |
| Input Modality | Audio (as of July 2026) |
| Output Modality | Text (as of July 2026) |
| Supported File Types | mp3, mp4, mpeg, mpga, m4a, wav, and webm (as of July 2026) |
| File Upload Limit | 25 MB per file for the OpenAI Transcriptions API (as of July 2026) |
| API Access | OpenAI Audio API and Gate.AI speech-to-text API (as of July 2026) |
| Translation Direction | Supported-language speech into English (as of July 2026) |
| Word-Level Timestamps | Supported through timestamp_granularities with whisper-1 (as of July 2026) |
| Knowledge Cutoff | Not applicable in the conventional LLM sense; not specified by OpenAI (as of July 2026) |
| Rate Limits | Account- and usage-tier dependent; no universal fixed rate is stated here (as of July 2026) |
| Fine-Tuning Support | Not confirmed for the hosted whisper-1 API (as of July 2026) |
| Native Realtime Support | Not specified for whisper-1; OpenAI documents separate realtime transcription models and APIs (as of July 2026) |
| Batch API Support | Not confirmed for whisper-1 in the reviewed documentation (as of July 2026) |
| Speaker Diarization | Not confirmed as a native whisper-1 capability (as of July 2026) |
| Tool or Function Calling | Not applicable to this speech-recognition model (as of July 2026) |
| General JSON Mode | Not applicable; transcription-specific response formats are available (as of July 2026) |
| Open-Source Availability | Whisper code and model weights are available separately from the hosted API (as of July 2026) |
| License and Usage Restrictions | Hosted API use follows OpenAI service terms; the open-source repository has its own license terms (as of July 2026) |
OpenAI’s model documentation lists Whisper 1 at \ $ 0.006 per minute and identifies audio as its input modality and text as its output modality. It does not publish separate input-token, cached-token, or output-token rates for this model.
As per the Gate.AI listing at https://gate.ai/models , conventional input and output token prices are not displayed for openai/stt-whisper-1 . Gate.AI’s API documentation states that whisper-1 is typically billed by audio duration, distinguishing it from token-billed transcription models.
What Can Whisper 1 Do That Makes It Useful in Production?
Multilingual audio transcription
Whisper 1 converts speech into text in the language spoken in the recording. This is useful for interviews, lectures, customer research, internal meetings, podcasts, and media archives. Accuracy still depends on the language, accent, microphone quality, background noise, speaker overlap, and subject-specific terminology.
Speech translation into English
OpenAI documents two principal speech-to-text workflows: transcribing audio in its original language and translating supported speech into English. This can reduce the number of processing stages needed for multilingual content review, although Whisper 1 does not provide arbitrary speech translation into every possible target language.
Word- and segment-level timestamps
The timestamp_granularities parameter can return timestamped output at the segment level, word level, or both. OpenAI specifically states that this parameter is supported for whisper-1 . Timestamp data is useful for subtitle alignment, transcript navigation, media editing, quotation verification, and searchable playback interfaces.
Common audio and media formats
The Transcriptions API accepts widely used formats, including MP3, MP4, MPEG, MPGA, M4A, WAV, and WebM. This reduces format-conversion work for many recorded-media pipelines, although developers may still need to compress or divide large files.
Prompt-guided transcription
The API can accept contextual prompt information to help establish expected terminology, names, acronyms, or writing style. A prompt may improve consistency, but it does not guarantee correct recognition of uncommon names or specialist vocabulary.
What Are Whisper 1’s Supported Modalities?
| Modality | Supported? | Notes |
|---|---|---|
| Audio input | Yes | Accepts supported recorded-audio and media-file formats |
| Text input | Limited | Optional instructions and transcription parameters; not general text generation |
| Image input | No | Whisper 1 does not process images |
| Video understanding | No | MP4 files may be accepted for their audio content, but visual frames are not analyzed |
| Text output | Yes | Produces transcription or English translation text |
| Audio output | No | Whisper 1 does not generate speech |
| Segment timestamps | Yes | Available through timestamped response formats |
| Word timestamps | Yes | Available through timestamp_granularities for whisper-1 |
| Native speaker labels | Not confirmed | Separate diarization processing may be required |
| Native realtime conversation | Not confirmed | Use a realtime transcription model when live streaming is required |
OpenAI currently limits Transcriptions API uploads to 25 MB and supports mp3 , mp4 , mpeg , mpga , m4a , wav , and webm . Files exceeding the limit must be compressed or divided into smaller chunks. OpenAI recommends avoiding cuts in the middle of sentences because chunk boundaries can remove useful context.
Where Does Whisper 1 Fall Short?
Whisper 1 is a specialized speech-recognition model. It does not independently summarize transcripts, answer questions about a recording, classify topics, generate images, synthesize speech, or perform general-purpose reasoning. Those tasks require another model or an application layer.
The 25 MB OpenAI upload limit can add preprocessing work for long interviews, lectures, meetings, podcasts, and video files. Chunking also creates a risk of lost context, inconsistent punctuation, repeated text, or broken sentences near segment boundaries.
Its translation workflow is designed to return English text. Applications that need speech translated directly into another non-English language generally require a separate translation stage.
Whisper 1 does not guarantee correct speaker attribution. Where speaker labels are important, teams should use a verified diarization capability or a separate speaker-segmentation system and test it on representative recordings.
Transcription quality can decline with overlapping speakers, distant microphones, music, heavy background noise, rapid code-switching, uncommon names, specialized terminology, or low-resource languages. Prompt context may help with expected vocabulary, but it should not be treated as a correctness guarantee.
This is a general AI limitation and is not unique to Whisper 1: generated transcripts may contain plausible but incorrect words, omissions, or punctuation. Legal, medical, financial, investigative, accessibility, and safety-critical transcripts require review against the original recording by an appropriately qualified person.
What Is Whisper 1 Best Used For?
| Use Case | Why Whisper 1 May Fit | Important Limitation |
|---|---|---|
| Interview transcription | Converts recorded speech into editable, searchable text | Names, quotations, and speaker changes require review |
| Meeting records | Creates a text record for search and downstream summarization | Native speaker diarization is not confirmed |
| Lecture and webinar transcription | Supports long-form spoken content when files are prepared correctly | Large recordings may require compression or chunking |
| Subtitle preparation | Provides timestamped output for alignment and editing | Timing and wording should be checked manually |
| Podcast and media indexing | Makes spoken content searchable and easier to catalog | Music and overlapping speakers can reduce accuracy |
| Multilingual research | Transcribes supported speech in its original language | Accuracy varies by language and recording quality |
| Speech-to-English translation | Produces English text from supported non-English audio | English is the documented translation target |
| Archive digitization | Converts recorded collections into searchable text | Poor or degraded source audio can reduce reliability |
| Content moderation support | Creates text that can be reviewed by downstream systems | Audio meaning and context may be lost in transcription |
| Accessibility workflows | Provides a draft transcript or caption source | Human correction is important before publication |
The phrase "best used for" is scenario-dependent. Whisper 1 may be suitable where a stable, file-based transcription API, English translation, or word-level timestamps are required. It should be evaluated on recordings that reflect the intended languages, accents, acoustic conditions, and vocabulary.
How Does Whisper 1 Compare to GPT-4o Transcribe and GPT-4o Mini Transcribe?
| Comparison Area | Whisper 1 | GPT-4o Transcribe | GPT-4o Mini Transcribe | Scenario Fit |
|---|---|---|---|---|
| Primary Purpose | General-purpose multilingual speech recognition | Newer OpenAI transcription model | Smaller newer OpenAI transcription model | All three address speech-to-text workflows |
| Audio Input | Yes | Yes | Yes | Relevant to recorded-audio transcription |
| Text Output | Yes | Yes | Yes | All return text transcripts |
| OpenAI Price | \$0.006 per audio minute | Check current OpenAI pricing before deployment | Check current OpenAI pricing before deployment | Compare total workload cost using current pricing |
| English Translation | Clearly documented for Whisper 1 | Confirm current endpoint support | Confirm current endpoint support | Whisper 1 may fit established speech-to-English workflows |
| Word-Level Timestamps | Explicitly documented for whisper-1 | Confirm current feature support | Confirm current feature support | Whisper 1 fits applications already built around timestamped output |
| API Maturity | Established hosted model with long-standing integrations | Newer model | Newer efficiency-oriented model | Existing systems may favor compatibility; new systems should benchmark alternatives |
| Native Diarization | Not confirmed | Check current model documentation | Check current model documentation | Speaker-labelled workflows require feature verification |
| Realtime Use | Not specified as a native realtime model | Check current realtime support | Check current realtime support | Live applications should use a model explicitly documented for realtime transcription |
No transcription model is universally preferable. Whisper 1 may fit established integrations that depend on its model ID, translation behavior, or timestamp format. New applications should test Whisper 1 and newer transcription models on representative audio, then compare accuracy, latency, response features, and total cost.
How Do I Access Whisper 1 Through Gate.AI?
As per the Gate.AI model-card at https://gate.ai/models , the Gate.AI model ID is:
openai/stt-whisper- 1
Gate.AI documents an OpenAI-compatible API base URL:
https://api.gate.ai/openai/v1
Its speech-to-text reference specifies POST /audio/transcriptions , Bearer-token authentication, a multipart/form-data request body, synchronous text output, usage information, and duration-based billing for whisper-1 .
Python Example
Python import os
from pathlib import Path
import OpenAI from openai
audio_path = Path("audio.mp3")
if audio_path.is_file():
raise FileNotFoundError(f"Audio file not found: {audio_path}")
client = OpenAI(
api_key=os.environ["GATEAI_API_KEY"],
base_url="https://api.gate.ai/openai/v1",
)
with audio_path.open("rb") as audio_file:
transcription = client.audio.transcriptions.create(
model="openai/stt-whisper-1",
file=audio_file,
)
print(transcription.text )
curl Example
``` Bash curl —request POST \
—url "https://api.gate.ai/openai/v1/audio/transcriptions" \
—header "Authorization: Bearer ${GATEAI_API_KEY}" \
—form "file=@audio.mp3" \
—form "model=openai/stt-whisper-1" ```
The examples keep the API key in the GATEAI_API_KEY environment variable rather than embedding it in source code. Applications should also validate file type, file size, error responses, retry behavior, and account limits before production deployment.
Gate.AI describes the speech-to-text endpoint as synchronous. Long or high-volume workloads should therefore be designed with request timeouts, queuing, file validation, and retry controls appropriate to the application.
Developers can also access the model directly through OpenAI’s /v1/audio/transcriptions and /v1/audio/translations APIs using the OpenAI model ID whisper-1 .
FAQs
What is Whisper 1’s context window?
OpenAI does not specify Whisper 1 with a conventional token context window because it is an audio transcription model rather than a chat LLM. The more relevant documented API constraint is a 25 MB file-upload limit for the OpenAI Transcriptions API.
How much does Whisper 1 cost?
OpenAI lists Whisper 1 at $ 0.006 per audio minute as of July 2026. It does not publish separate input-token, cached-input, or output-token prices. Gate.AI documents duration-based billing for whisper-1 ; current account pricing should be checked before deployment.
How can developers access Whisper 1?
Developers can access it directly through OpenAI’s Audio API with the model ID whisper-1 . As per the Gate.AI model-card, it is also available as openai/stt-whisper-1 through Gate.AI’s OpenAI-compatible /audio/transcriptions endpoint.
What is Whisper 1 suitable for?
Whisper 1 may fit recorded interviews, meetings, lectures, subtitles, podcasts, multilingual transcription, searchable audio archives, and speech-to-English translation. Accuracy varies with language, accent, recording quality, overlapping speech, background noise, and specialist vocabulary, so important transcripts require human review.
